## check links library("cotram") library("survival") set.seed(29) ### check coefficients (tram vs cotram model) .check_cf <- function(m, mc, ...) { cfc <- coef(as.mlt(mc)) ## tram coefs for negative = TRUE cf <- coef(c <- as.mlt(m)) cf[c$shiftcoef] <- c(-1, 1)[(m$negative & mc$negative) + 1] * cf[c$shiftcoef] ## check tram vs cotram coefs stopifnot(all.equal(cf, cfc, check.attributes = FALSE, ...)) } ### check log-likelihood (tram vs cotram model) .check_ll <- function(m, mc, ...) { ## simple check wrt to newdata stopifnot(logLik(mc) == logLik(mc, newdata = model.frame(mc))) ## likelihood contributions interval-censored ### newdata nd_m <- d[c("y", "x")] colnames(nd_m) <- colnames(m$data) nd_m1 <- nd_m2 <- nd_m nd_m2[, 1] <- nd_m2[, 1] + as.integer(log_first) nd_m1[, 1] <- (nd_m1[, 1] - 1L) + as.integer(log_first) L <- predict(m, newdata = nd_m2, type = "distribution") - predict(m, newdata = nd_m1, type = "distribution") ## check interval-censored vs cotram log-likelihood contributions stopifnot(all.equal(log(L), mc$logliki(coef(as.mlt(mc)), mc$weights), check.attributes = FALSE, ...)) # check tram vs cotram log-likelihood contributions stopifnot(all.equal(m$logliki(coef(as.mlt(m)), m$weight), mc$logliki(coef(as.mlt(mc)), mc$weight), ...)) } ### run checks for log_first = FALSE / TRUE & all links ## dgp counts n <- 200 x <- runif(n) y <- as.integer(rnbinom(n, mu = exp(.5 + .8 * x), size = 10)) ## trams & cotrams links <- c("logit", "cloglog", "loglog", "probit") trams <- c("logit" = "Colr", "cloglog" = "Coxph", "loglog" = "Lehmann", "probit" = "BoxCox") ## run for (log_first in c(FALSE, TRUE)) { set.seed(2) # print(log_first) ## plus_one for log_first = TRUE plus_one <- as.integer(log_first) ## plus_one for integer y yii <- y + plus_one ## counts as interval censored object yleft <- y - 1L yleft[yleft < 0] <- -Inf yi <- Surv(yleft + plus_one, y + plus_one, type = "interval2") yi2 <- cotram:::.R.count(as.integer(y), plus_one = log_first) all.equal(yi, yi2) d <- data.frame(y = y, yi = yi, x = x) for (link in links) { # print(link) ## fit "cotram" mc <- cotram(as.formula(y ~ x), data = d, method = link, log_first = log_first) ## check model.frame stopifnot(all.equal(nd <- model.frame(mc), d[names(nd)], check.attributes = FALSE)) ## "tram" with interval-censored response tram <- unname(trams[link]) m <- do.call(tram, list(formula = yi ~ x, support = mc$support, bounds = mc$bounds, log_first = log_first)) ## check if same model stopifnot(mc$model$todistr$name == m$model$todistr$name) ## check coefs and logLiks .check_cf(m, mc) .check_ll(m, mc) ## "tram" with integer response m2 <- do.call(tram, list(formula = yii ~ x, support = c(mc$support), bounds = mc$bounds, log_first = log_first)) ## check if same model stopifnot(mc$model$todistr$name == m2$model$todistr$name) ## check coefs and logLiks ### minor differences because integer are treated as interval (y - 1, 1] in "tram" .check_cf(m2, mc, tol = 1e-6) .check_ll(m2, mc, tol = 1e-6) } } ## NOTE: ## - Models with integer-y and interval-censored-y (for [y - 1, y]) are not ## exactly equivalent for "tram": This is because the integer is treated as interval (y - 1, y] ## => This also explains the discrepancies between "cotram" and integer-y with "tram" ## - The discrepancy between "cotram" and interval-censored "tram" fit with "prob" given, ## comes from the different support (is rounded it "cotram") ## additional checks for plus_one d0 <- data.frame(y = q <- 0L:100L, x = runif(length(q))) m0 <- cotram(y ~ x, log_first = TRUE, data = d0) stopifnot(all(mkgrid(m0)$y == q)) ## mkgrid for log_first m0 <- cotram(y ~ x, log_first = FALSE, data = d0) stopifnot(all(mkgrid(m0)$y == q)) ## mkgrid